Fix null pointer crash with adjustAndMark

Source/platform/heap/Visitor.h

 /*
  * Copyright (C) 2013 Google Inc. All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are
  * met:
  *
  *     * Redistributions of source code must retain the above copyright
  * notice, this list of conditions and the following disclaimer.
  *     * Redistributions in binary form must reproduce the above
@@ skipping 202 matching lines @@
 public:
     virtual ~Visitor() { }
 
     // One-argument templated mark method. This uses the static type of
     // the argument to get the TraceTrait. By default, the mark method
     // of the TraceTrait just calls the virtual two-argument mark method on this
     // visitor, where the second argument is the static trace method of the trait.
     template<typename T>
     void mark(T* t)
     {
         if (!t)

[haraken, 2014/06/13 09:14:16]

I was thinking that this null check is enough. Isn't this mark() method called
before we hit self->adjustAndMark(visitor)?

[Erik Corry, 2014/06/13 09:40:26]

Not for members in vector backing.

             return;
 #ifndef NDEBUG
         TraceTrait<T>::checkGCInfo(this, t);
 #endif
         TraceTrait<T>::mark(this, t);
     }
 
     // Member version of the one-argument templated trace method.
     template<typename T>
     void trace(const Member<T>& t)
@@ skipping 380 matching lines @@
         visitor->checkGCInfo(const_cast<T*>(t), GCInfoTrait<T>::get());
     }
 #endif
 };
 
 template<typename T>
 class DefaultTraceTrait<T, true> {
 public:
     static void mark(Visitor* visitor, const T* self)
     {
-        self->adjustAndMark(visitor);
+        if (self)

[sof, 2014/06/13 09:12:54]

Would it be possible move this "up" a bit? i.e., we're paying for a null check
on pointers that are otherwise guaranteed to be non-null now, right?

[Erik Corry, 2014/06/13 09:40:26]

This is just for things that need adjustAndMark, so I'm not too worried about
the performance issues.  Also I think once stuff has been inlined any extra
checks are pretty simple to optimize out.

[sof, 2014/06/13 09:49:35]

alright, perhaps so.

+            self->adjustAndMark(visitor);
     }
 
 #ifndef NDEBUG
     static void checkGCInfo(Visitor*, const T*) { }
 #endif
 };
 
 template<typename T, bool = NeedsAdjustAndMark<T>::value> class DefaultObjectAliveTrait;
 
 template<typename T>
@@ skipping 111 matching lines @@
 struct GCInfoTrait {
     static const GCInfo* get()
     {
         return GCInfoAtBase<typename GetGarbageCollectedBase<T>::type>::get();
     }
 };
 
 }
 
 #endif